3 * Copyright (c) 2016 Timo Rothenpieler <timo@rothenpieler.org>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "compat/cuda/dynlink_loader.h"
24 #include "libavutil/buffer.h"
25 #include "libavutil/mathematics.h"
26 #include "libavutil/hwcontext.h"
27 #include "libavutil/hwcontext_cuda_internal.h"
28 #include "libavutil/cuda_check.h"
29 #include "libavutil/fifo.h"
30 #include "libavutil/log.h"
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
40 #if !NVDECAPI_CHECK_VERSION(9, 0)
41 #define cudaVideoSurfaceFormat_YUV444 2
42 #define cudaVideoSurfaceFormat_YUV444_16Bit 3
45 typedef struct CuvidContext
49 CUvideodecoder cudecoder;
50 CUvideoparser cuparser;
54 int drop_second_field;
70 AVBufferRef *hwdevice;
75 AVFifoBuffer *frame_queue;
78 int deint_mode_current;
80 int progressive_sequence;
87 cudaVideoCodec codec_type;
88 cudaVideoChromaFormat chroma_format;
90 CUVIDDECODECAPS caps8, caps10, caps12;
92 CUVIDPARSERPARAMS cuparseinfo;
93 CUVIDEOFORMATEX cuparse_ext;
99 typedef struct CuvidParsedFrame
101 CUVIDPARSERDISPINFO dispinfo;
103 int is_deinterlacing;
106 #define CHECK_CU(x) FF_CUDA_CHECK_DL(avctx, ctx->cudl, x)
108 static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
110 AVCodecContext *avctx = opaque;
111 CuvidContext *ctx = avctx->priv_data;
112 AVHWFramesContext *hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
113 CUVIDDECODECAPS *caps = NULL;
114 CUVIDDECODECREATEINFO cuinfo;
118 int old_width = avctx->width;
119 int old_height = avctx->height;
121 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
122 AV_PIX_FMT_NONE, // Will be updated below
125 av_log(avctx, AV_LOG_TRACE, "pfnSequenceCallback, progressive_sequence=%d\n", format->progressive_sequence);
127 memset(&cuinfo, 0, sizeof(cuinfo));
129 ctx->internal_error = 0;
131 avctx->coded_width = cuinfo.ulWidth = format->coded_width;
132 avctx->coded_height = cuinfo.ulHeight = format->coded_height;
135 cuinfo.display_area.left = format->display_area.left + ctx->crop.left;
136 cuinfo.display_area.top = format->display_area.top + ctx->crop.top;
137 cuinfo.display_area.right = format->display_area.right - ctx->crop.right;
138 cuinfo.display_area.bottom = format->display_area.bottom - ctx->crop.bottom;
140 // width and height need to be set before calling ff_get_format
141 if (ctx->resize_expr) {
142 avctx->width = ctx->resize.width;
143 avctx->height = ctx->resize.height;
145 avctx->width = cuinfo.display_area.right - cuinfo.display_area.left;
146 avctx->height = cuinfo.display_area.bottom - cuinfo.display_area.top;
149 // target width/height need to be multiples of two
150 cuinfo.ulTargetWidth = avctx->width = (avctx->width + 1) & ~1;
151 cuinfo.ulTargetHeight = avctx->height = (avctx->height + 1) & ~1;
153 // aspect ratio conversion, 1:1, depends on scaled resolution
154 cuinfo.target_rect.left = 0;
155 cuinfo.target_rect.top = 0;
156 cuinfo.target_rect.right = cuinfo.ulTargetWidth;
157 cuinfo.target_rect.bottom = cuinfo.ulTargetHeight;
159 chroma_444 = format->chroma_format == cudaVideoChromaFormat_444;
161 switch (format->bit_depth_luma_minus8) {
163 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P : AV_PIX_FMT_NV12;
167 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P010;
171 pix_fmts[1] = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P016;
178 if (!caps || !caps->bIsSupported) {
179 av_log(avctx, AV_LOG_ERROR, "unsupported bit depth: %d\n",
180 format->bit_depth_luma_minus8 + 8);
181 ctx->internal_error = AVERROR(EINVAL);
185 surface_fmt = ff_get_format(avctx, pix_fmts);
186 if (surface_fmt < 0) {
187 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", surface_fmt);
188 ctx->internal_error = AVERROR(EINVAL);
192 av_log(avctx, AV_LOG_VERBOSE, "Formats: Original: %s | HW: %s | SW: %s\n",
193 av_get_pix_fmt_name(avctx->pix_fmt),
194 av_get_pix_fmt_name(surface_fmt),
195 av_get_pix_fmt_name(avctx->sw_pix_fmt));
197 avctx->pix_fmt = surface_fmt;
199 // Update our hwframe ctx, as the get_format callback might have refreshed it!
200 if (avctx->hw_frames_ctx) {
201 av_buffer_unref(&ctx->hwframe);
203 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
205 ctx->internal_error = AVERROR(ENOMEM);
209 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
212 ff_set_sar(avctx, av_div_q(
213 (AVRational){ format->display_aspect_ratio.x, format->display_aspect_ratio.y },
214 (AVRational){ avctx->width, avctx->height }));
216 ctx->deint_mode_current = format->progressive_sequence
217 ? cudaVideoDeinterlaceMode_Weave
220 ctx->progressive_sequence = format->progressive_sequence;
222 if (!format->progressive_sequence && ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave)
223 avctx->flags |= AV_CODEC_FLAG_INTERLACED_DCT;
225 avctx->flags &= ~AV_CODEC_FLAG_INTERLACED_DCT;
227 if (format->video_signal_description.video_full_range_flag)
228 avctx->color_range = AVCOL_RANGE_JPEG;
230 avctx->color_range = AVCOL_RANGE_MPEG;
232 avctx->color_primaries = format->video_signal_description.color_primaries;
233 avctx->color_trc = format->video_signal_description.transfer_characteristics;
234 avctx->colorspace = format->video_signal_description.matrix_coefficients;
237 avctx->bit_rate = format->bitrate;
239 if (format->frame_rate.numerator && format->frame_rate.denominator) {
240 avctx->framerate.num = format->frame_rate.numerator;
241 avctx->framerate.den = format->frame_rate.denominator;
245 && avctx->coded_width == format->coded_width
246 && avctx->coded_height == format->coded_height
247 && avctx->width == old_width
248 && avctx->height == old_height
249 && ctx->chroma_format == format->chroma_format
250 && ctx->codec_type == format->codec)
253 if (ctx->cudecoder) {
254 av_log(avctx, AV_LOG_TRACE, "Re-initializing decoder\n");
255 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder));
256 if (ctx->internal_error < 0)
258 ctx->cudecoder = NULL;
261 if (hwframe_ctx->pool && (
262 hwframe_ctx->width < avctx->width ||
263 hwframe_ctx->height < avctx->height ||
264 hwframe_ctx->format != AV_PIX_FMT_CUDA ||
265 hwframe_ctx->sw_format != avctx->sw_pix_fmt)) {
266 av_log(avctx, AV_LOG_ERROR, "AVHWFramesContext is already initialized with incompatible parameters\n");
267 av_log(avctx, AV_LOG_DEBUG, "width: %d <-> %d\n", hwframe_ctx->width, avctx->width);
268 av_log(avctx, AV_LOG_DEBUG, "height: %d <-> %d\n", hwframe_ctx->height, avctx->height);
269 av_log(avctx, AV_LOG_DEBUG, "format: %s <-> cuda\n", av_get_pix_fmt_name(hwframe_ctx->format));
270 av_log(avctx, AV_LOG_DEBUG, "sw_format: %s <-> %s\n",
271 av_get_pix_fmt_name(hwframe_ctx->sw_format), av_get_pix_fmt_name(avctx->sw_pix_fmt));
272 ctx->internal_error = AVERROR(EINVAL);
276 ctx->chroma_format = format->chroma_format;
278 cuinfo.CodecType = ctx->codec_type = format->codec;
279 cuinfo.ChromaFormat = format->chroma_format;
281 switch (avctx->sw_pix_fmt) {
282 case AV_PIX_FMT_NV12:
283 cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
285 case AV_PIX_FMT_P010:
286 case AV_PIX_FMT_P016:
287 cuinfo.OutputFormat = cudaVideoSurfaceFormat_P016;
289 case AV_PIX_FMT_YUV444P:
290 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444;
292 case AV_PIX_FMT_YUV444P16:
293 cuinfo.OutputFormat = cudaVideoSurfaceFormat_YUV444_16Bit;
296 av_log(avctx, AV_LOG_ERROR, "Unsupported output format: %s\n",
297 av_get_pix_fmt_name(avctx->sw_pix_fmt));
298 ctx->internal_error = AVERROR(EINVAL);
302 cuinfo.ulNumDecodeSurfaces = ctx->nb_surfaces;
303 cuinfo.ulNumOutputSurfaces = 1;
304 cuinfo.ulCreationFlags = cudaVideoCreate_PreferCUVID;
305 cuinfo.bitDepthMinus8 = format->bit_depth_luma_minus8;
306 cuinfo.DeinterlaceMode = ctx->deint_mode_current;
308 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
309 avctx->framerate = av_mul_q(avctx->framerate, (AVRational){2, 1});
311 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidCreateDecoder(&ctx->cudecoder, &cuinfo));
312 if (ctx->internal_error < 0)
315 if (!hwframe_ctx->pool) {
316 hwframe_ctx->format = AV_PIX_FMT_CUDA;
317 hwframe_ctx->sw_format = avctx->sw_pix_fmt;
318 hwframe_ctx->width = avctx->width;
319 hwframe_ctx->height = avctx->height;
321 if ((ctx->internal_error = av_hwframe_ctx_init(ctx->hwframe)) < 0) {
322 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_init failed\n");
330 static int CUDAAPI cuvid_handle_picture_decode(void *opaque, CUVIDPICPARAMS* picparams)
332 AVCodecContext *avctx = opaque;
333 CuvidContext *ctx = avctx->priv_data;
335 av_log(avctx, AV_LOG_TRACE, "pfnDecodePicture\n");
337 ctx->key_frame[picparams->CurrPicIdx] = picparams->intra_pic_flag;
339 ctx->internal_error = CHECK_CU(ctx->cvdl->cuvidDecodePicture(ctx->cudecoder, picparams));
340 if (ctx->internal_error < 0)
346 static int CUDAAPI cuvid_handle_picture_display(void *opaque, CUVIDPARSERDISPINFO* dispinfo)
348 AVCodecContext *avctx = opaque;
349 CuvidContext *ctx = avctx->priv_data;
350 CuvidParsedFrame parsed_frame = { { 0 } };
352 parsed_frame.dispinfo = *dispinfo;
353 ctx->internal_error = 0;
355 // For some reason, dispinfo->progressive_frame is sometimes wrong.
356 parsed_frame.dispinfo.progressive_frame = ctx->progressive_sequence;
358 if (ctx->deint_mode_current == cudaVideoDeinterlaceMode_Weave) {
359 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
361 parsed_frame.is_deinterlacing = 1;
362 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
363 if (!ctx->drop_second_field) {
364 parsed_frame.second_field = 1;
365 av_fifo_generic_write(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
372 static int cuvid_is_buffer_full(AVCodecContext *avctx)
374 CuvidContext *ctx = avctx->priv_data;
376 int delay = ctx->cuparseinfo.ulMaxDisplayDelay;
377 if (ctx->deint_mode != cudaVideoDeinterlaceMode_Weave && !ctx->drop_second_field)
380 return (av_fifo_size(ctx->frame_queue) / sizeof(CuvidParsedFrame)) + delay >= ctx->nb_surfaces;
383 static int cuvid_decode_packet(AVCodecContext *avctx, const AVPacket *avpkt)
385 CuvidContext *ctx = avctx->priv_data;
386 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
387 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
388 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
389 CUVIDSOURCEDATAPACKET cupkt;
390 AVPacket filter_packet = { 0 };
391 AVPacket filtered_packet = { 0 };
392 int ret = 0, eret = 0, is_flush = ctx->decoder_flushing;
394 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_packet\n");
396 if (is_flush && avpkt && avpkt->size)
399 if (cuvid_is_buffer_full(avctx) && avpkt && avpkt->size)
400 return AVERROR(EAGAIN);
402 if (ctx->bsf && avpkt && avpkt->size) {
403 if ((ret = av_packet_ref(&filter_packet, avpkt)) < 0) {
404 av_log(avctx, AV_LOG_ERROR, "av_packet_ref failed\n");
408 if ((ret = av_bsf_send_packet(ctx->bsf, &filter_packet)) < 0) {
409 av_log(avctx, AV_LOG_ERROR, "av_bsf_send_packet failed\n");
410 av_packet_unref(&filter_packet);
414 if ((ret = av_bsf_receive_packet(ctx->bsf, &filtered_packet)) < 0) {
415 av_log(avctx, AV_LOG_ERROR, "av_bsf_receive_packet failed\n");
419 avpkt = &filtered_packet;
422 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
424 av_packet_unref(&filtered_packet);
428 memset(&cupkt, 0, sizeof(cupkt));
430 if (avpkt && avpkt->size) {
431 cupkt.payload_size = avpkt->size;
432 cupkt.payload = avpkt->data;
434 if (avpkt->pts != AV_NOPTS_VALUE) {
435 cupkt.flags = CUVID_PKT_TIMESTAMP;
436 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
437 cupkt.timestamp = av_rescale_q(avpkt->pts, avctx->pkt_timebase, (AVRational){1, 10000000});
439 cupkt.timestamp = avpkt->pts;
442 cupkt.flags = CUVID_PKT_ENDOFSTREAM;
443 ctx->decoder_flushing = 1;
446 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &cupkt));
448 av_packet_unref(&filtered_packet);
453 // cuvidParseVideoData doesn't return an error just because stuff failed...
454 if (ctx->internal_error) {
455 av_log(avctx, AV_LOG_ERROR, "cuvid decode callback error\n");
456 ret = ctx->internal_error;
461 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
473 static int cuvid_output_frame(AVCodecContext *avctx, AVFrame *frame)
475 CuvidContext *ctx = avctx->priv_data;
476 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
477 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
478 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
479 CUdeviceptr mapped_frame = 0;
480 int ret = 0, eret = 0;
482 av_log(avctx, AV_LOG_TRACE, "cuvid_output_frame\n");
484 if (ctx->decoder_flushing) {
485 ret = cuvid_decode_packet(avctx, NULL);
486 if (ret < 0 && ret != AVERROR_EOF)
490 if (!cuvid_is_buffer_full(avctx)) {
492 ret = ff_decode_get_packet(avctx, &pkt);
493 if (ret < 0 && ret != AVERROR_EOF)
495 ret = cuvid_decode_packet(avctx, &pkt);
496 av_packet_unref(&pkt);
497 // cuvid_is_buffer_full() should avoid this.
498 if (ret == AVERROR(EAGAIN))
499 ret = AVERROR_EXTERNAL;
500 if (ret < 0 && ret != AVERROR_EOF)
504 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
508 if (av_fifo_size(ctx->frame_queue)) {
509 const AVPixFmtDescriptor *pixdesc;
510 CuvidParsedFrame parsed_frame;
511 CUVIDPROCPARAMS params;
512 unsigned int pitch = 0;
516 av_fifo_generic_read(ctx->frame_queue, &parsed_frame, sizeof(CuvidParsedFrame), NULL);
518 memset(¶ms, 0, sizeof(params));
519 params.progressive_frame = parsed_frame.dispinfo.progressive_frame;
520 params.second_field = parsed_frame.second_field;
521 params.top_field_first = parsed_frame.dispinfo.top_field_first;
523 ret = CHECK_CU(ctx->cvdl->cuvidMapVideoFrame(ctx->cudecoder, parsed_frame.dispinfo.picture_index, &mapped_frame, &pitch, ¶ms));
527 if (avctx->pix_fmt == AV_PIX_FMT_CUDA) {
528 ret = av_hwframe_get_buffer(ctx->hwframe, frame, 0);
530 av_log(avctx, AV_LOG_ERROR, "av_hwframe_get_buffer failed\n");
534 ret = ff_decode_frame_props(avctx, frame);
536 av_log(avctx, AV_LOG_ERROR, "ff_decode_frame_props failed\n");
540 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
542 for (i = 0; i < pixdesc->nb_components; i++) {
543 int height = avctx->height >> (i ? pixdesc->log2_chroma_h : 0);
544 CUDA_MEMCPY2D cpy = {
545 .srcMemoryType = CU_MEMORYTYPE_DEVICE,
546 .dstMemoryType = CU_MEMORYTYPE_DEVICE,
547 .srcDevice = mapped_frame,
548 .dstDevice = (CUdeviceptr)frame->data[i],
550 .dstPitch = frame->linesize[i],
552 .WidthInBytes = FFMIN(pitch, frame->linesize[i]),
556 ret = CHECK_CU(ctx->cudl->cuMemcpy2DAsync(&cpy, device_hwctx->stream));
562 } else if (avctx->pix_fmt == AV_PIX_FMT_NV12 ||
563 avctx->pix_fmt == AV_PIX_FMT_P010 ||
564 avctx->pix_fmt == AV_PIX_FMT_P016 ||
565 avctx->pix_fmt == AV_PIX_FMT_YUV444P ||
566 avctx->pix_fmt == AV_PIX_FMT_YUV444P16) {
567 unsigned int offset = 0;
568 AVFrame *tmp_frame = av_frame_alloc();
570 av_log(avctx, AV_LOG_ERROR, "av_frame_alloc failed\n");
571 ret = AVERROR(ENOMEM);
575 pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
577 tmp_frame->format = AV_PIX_FMT_CUDA;
578 tmp_frame->hw_frames_ctx = av_buffer_ref(ctx->hwframe);
579 tmp_frame->width = avctx->width;
580 tmp_frame->height = avctx->height;
583 * Note that the following logic would not work for three plane
584 * YUV420 because the pitch value is different for the chroma
587 for (i = 0; i < pixdesc->nb_components; i++) {
588 tmp_frame->data[i] = (uint8_t*)mapped_frame + offset;
589 tmp_frame->linesize[i] = pitch;
590 offset += pitch * (avctx->height >> (i ? pixdesc->log2_chroma_h : 0));
593 ret = ff_get_buffer(avctx, frame, 0);
595 av_log(avctx, AV_LOG_ERROR, "ff_get_buffer failed\n");
596 av_frame_free(&tmp_frame);
600 ret = av_hwframe_transfer_data(frame, tmp_frame, 0);
602 av_log(avctx, AV_LOG_ERROR, "av_hwframe_transfer_data failed\n");
603 av_frame_free(&tmp_frame);
606 av_frame_free(&tmp_frame);
612 frame->key_frame = ctx->key_frame[parsed_frame.dispinfo.picture_index];
613 frame->width = avctx->width;
614 frame->height = avctx->height;
615 if (avctx->pkt_timebase.num && avctx->pkt_timebase.den)
616 frame->pts = av_rescale_q(parsed_frame.dispinfo.timestamp, (AVRational){1, 10000000}, avctx->pkt_timebase);
618 frame->pts = parsed_frame.dispinfo.timestamp;
620 if (parsed_frame.second_field) {
621 if (ctx->prev_pts == INT64_MIN) {
622 ctx->prev_pts = frame->pts;
623 frame->pts += (avctx->pkt_timebase.den * avctx->framerate.den) / (avctx->pkt_timebase.num * avctx->framerate.num);
625 int pts_diff = (frame->pts - ctx->prev_pts) / 2;
626 ctx->prev_pts = frame->pts;
627 frame->pts += pts_diff;
631 /* CUVIDs opaque reordering breaks the internal pkt logic.
632 * So set pkt_pts and clear all the other pkt_ fields.
635 FF_DISABLE_DEPRECATION_WARNINGS
636 frame->pkt_pts = frame->pts;
637 FF_ENABLE_DEPRECATION_WARNINGS
640 frame->pkt_duration = 0;
641 frame->pkt_size = -1;
643 frame->interlaced_frame = !parsed_frame.is_deinterlacing && !parsed_frame.dispinfo.progressive_frame;
645 if (frame->interlaced_frame)
646 frame->top_field_first = parsed_frame.dispinfo.top_field_first;
647 } else if (ctx->decoder_flushing) {
650 ret = AVERROR(EAGAIN);
655 eret = CHECK_CU(ctx->cvdl->cuvidUnmapVideoFrame(ctx->cudecoder, mapped_frame));
657 eret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
665 static int cuvid_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt)
667 CuvidContext *ctx = avctx->priv_data;
668 AVFrame *frame = data;
671 av_log(avctx, AV_LOG_TRACE, "cuvid_decode_frame\n");
673 if (ctx->deint_mode_current != cudaVideoDeinterlaceMode_Weave) {
674 av_log(avctx, AV_LOG_ERROR, "Deinterlacing is not supported via the old API\n");
675 return AVERROR(EINVAL);
678 if (!ctx->decoder_flushing) {
679 ret = cuvid_decode_packet(avctx, avpkt);
684 ret = cuvid_output_frame(avctx, frame);
685 if (ret == AVERROR(EAGAIN) || ret == AVERROR_EOF) {
687 } else if (ret < 0) {
696 static av_cold int cuvid_decode_end(AVCodecContext *avctx)
698 CuvidContext *ctx = avctx->priv_data;
700 av_fifo_freep(&ctx->frame_queue);
703 av_bsf_free(&ctx->bsf);
706 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
709 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
713 av_buffer_unref(&ctx->hwframe);
714 av_buffer_unref(&ctx->hwdevice);
716 av_freep(&ctx->key_frame);
718 cuvid_free_functions(&ctx->cvdl);
723 static int cuvid_test_capabilities(AVCodecContext *avctx,
724 const CUVIDPARSERPARAMS *cuparseinfo,
729 CuvidContext *ctx = avctx->priv_data;
730 CUVIDDECODECAPS *caps;
731 int res8 = 0, res10 = 0, res12 = 0;
733 if (!ctx->cvdl->cuvidGetDecoderCaps) {
734 av_log(avctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
735 av_log(avctx, AV_LOG_WARNING, "The minimum required version is "
736 #if defined(_WIN32) || defined(__CYGWIN__)
741 ". Continuing blind.\n");
742 ctx->caps8.bIsSupported = ctx->caps10.bIsSupported = 1;
743 // 12 bit was not supported before the capability check was introduced, so disable it.
744 ctx->caps12.bIsSupported = 0;
748 ctx->caps8.eCodecType = ctx->caps10.eCodecType = ctx->caps12.eCodecType
749 = cuparseinfo->CodecType;
750 ctx->caps8.eChromaFormat = ctx->caps10.eChromaFormat = ctx->caps12.eChromaFormat
751 = cudaVideoChromaFormat_420;
753 ctx->caps8.nBitDepthMinus8 = 0;
754 ctx->caps10.nBitDepthMinus8 = 2;
755 ctx->caps12.nBitDepthMinus8 = 4;
757 res8 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8));
758 res10 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10));
759 res12 = CHECK_CU(ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12));
761 av_log(avctx, AV_LOG_VERBOSE, "CUVID capabilities for %s:\n", avctx->codec->name);
762 av_log(avctx, AV_LOG_VERBOSE, "8 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
763 ctx->caps8.bIsSupported, ctx->caps8.nMinWidth, ctx->caps8.nMaxWidth, ctx->caps8.nMinHeight, ctx->caps8.nMaxHeight);
764 av_log(avctx, AV_LOG_VERBOSE, "10 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
765 ctx->caps10.bIsSupported, ctx->caps10.nMinWidth, ctx->caps10.nMaxWidth, ctx->caps10.nMinHeight, ctx->caps10.nMaxHeight);
766 av_log(avctx, AV_LOG_VERBOSE, "12 bit: supported: %d, min_width: %d, max_width: %d, min_height: %d, max_height: %d\n",
767 ctx->caps12.bIsSupported, ctx->caps12.nMinWidth, ctx->caps12.nMaxWidth, ctx->caps12.nMinHeight, ctx->caps12.nMaxHeight);
786 if (!ctx->caps8.bIsSupported) {
787 av_log(avctx, AV_LOG_ERROR, "Codec %s is not supported.\n", avctx->codec->name);
788 return AVERROR(EINVAL);
791 if (!caps->bIsSupported) {
792 av_log(avctx, AV_LOG_ERROR, "Bit depth %d is not supported.\n", bit_depth);
793 return AVERROR(EINVAL);
796 if (probed_width > caps->nMaxWidth || probed_width < caps->nMinWidth) {
797 av_log(avctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
798 probed_width, caps->nMinWidth, caps->nMaxWidth);
799 return AVERROR(EINVAL);
802 if (probed_height > caps->nMaxHeight || probed_height < caps->nMinHeight) {
803 av_log(avctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
804 probed_height, caps->nMinHeight, caps->nMaxHeight);
805 return AVERROR(EINVAL);
808 if ((probed_width * probed_height) / 256 > caps->nMaxMBCount) {
809 av_log(avctx, AV_LOG_ERROR, "Video macroblock count %d exceeds maximum of %d\n",
810 (int)(probed_width * probed_height) / 256, caps->nMaxMBCount);
811 return AVERROR(EINVAL);
817 static av_cold int cuvid_decode_init(AVCodecContext *avctx)
819 CuvidContext *ctx = avctx->priv_data;
820 AVCUDADeviceContext *device_hwctx;
821 AVHWDeviceContext *device_ctx;
822 AVHWFramesContext *hwframe_ctx;
823 CUVIDSOURCEDATAPACKET seq_pkt;
824 CUcontext cuda_ctx = NULL;
826 const AVBitStreamFilter *bsf;
829 enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
833 int probed_width = avctx->coded_width ? avctx->coded_width : 1280;
834 int probed_height = avctx->coded_height ? avctx->coded_height : 720;
835 int probed_bit_depth = 8;
837 const AVPixFmtDescriptor *probe_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
838 if (probe_desc && probe_desc->nb_components)
839 probed_bit_depth = probe_desc->comp[0].depth;
841 // Accelerated transcoding scenarios with 'ffmpeg' require that the
842 // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
843 // pix_fmt for non-accelerated transcoding, do not need to be correct
844 // but need to be set to something. We arbitrarily pick NV12.
845 ret = ff_get_format(avctx, pix_fmts);
847 av_log(avctx, AV_LOG_ERROR, "ff_get_format failed: %d\n", ret);
850 avctx->pix_fmt = ret;
852 if (ctx->resize_expr && sscanf(ctx->resize_expr, "%dx%d",
853 &ctx->resize.width, &ctx->resize.height) != 2) {
854 av_log(avctx, AV_LOG_ERROR, "Invalid resize expressions\n");
855 ret = AVERROR(EINVAL);
859 if (ctx->crop_expr && sscanf(ctx->crop_expr, "%dx%dx%dx%d",
860 &ctx->crop.top, &ctx->crop.bottom,
861 &ctx->crop.left, &ctx->crop.right) != 4) {
862 av_log(avctx, AV_LOG_ERROR, "Invalid cropping expressions\n");
863 ret = AVERROR(EINVAL);
867 ret = cuvid_load_functions(&ctx->cvdl, avctx);
869 av_log(avctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
873 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
874 if (!ctx->frame_queue) {
875 ret = AVERROR(ENOMEM);
879 if (avctx->hw_frames_ctx) {
880 ctx->hwframe = av_buffer_ref(avctx->hw_frames_ctx);
882 ret = AVERROR(ENOMEM);
886 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
888 ctx->hwdevice = av_buffer_ref(hwframe_ctx->device_ref);
889 if (!ctx->hwdevice) {
890 ret = AVERROR(ENOMEM);
894 if (avctx->hw_device_ctx) {
895 ctx->hwdevice = av_buffer_ref(avctx->hw_device_ctx);
896 if (!ctx->hwdevice) {
897 ret = AVERROR(ENOMEM);
901 ret = av_hwdevice_ctx_create(&ctx->hwdevice, AV_HWDEVICE_TYPE_CUDA, ctx->cu_gpu, NULL, 0);
906 ctx->hwframe = av_hwframe_ctx_alloc(ctx->hwdevice);
908 av_log(avctx, AV_LOG_ERROR, "av_hwframe_ctx_alloc failed\n");
909 ret = AVERROR(ENOMEM);
913 hwframe_ctx = (AVHWFramesContext*)ctx->hwframe->data;
916 device_ctx = hwframe_ctx->device_ctx;
917 device_hwctx = device_ctx->hwctx;
919 cuda_ctx = device_hwctx->cuda_ctx;
920 ctx->cudl = device_hwctx->internal->cuda_dl;
922 memset(&ctx->cuparseinfo, 0, sizeof(ctx->cuparseinfo));
923 memset(&ctx->cuparse_ext, 0, sizeof(ctx->cuparse_ext));
924 memset(&seq_pkt, 0, sizeof(seq_pkt));
926 ctx->cuparseinfo.pExtVideoInfo = &ctx->cuparse_ext;
928 switch (avctx->codec->id) {
929 #if CONFIG_H264_CUVID_DECODER
930 case AV_CODEC_ID_H264:
931 ctx->cuparseinfo.CodecType = cudaVideoCodec_H264;
934 #if CONFIG_HEVC_CUVID_DECODER
935 case AV_CODEC_ID_HEVC:
936 ctx->cuparseinfo.CodecType = cudaVideoCodec_HEVC;
939 #if CONFIG_MJPEG_CUVID_DECODER
940 case AV_CODEC_ID_MJPEG:
941 ctx->cuparseinfo.CodecType = cudaVideoCodec_JPEG;
944 #if CONFIG_MPEG1_CUVID_DECODER
945 case AV_CODEC_ID_MPEG1VIDEO:
946 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG1;
949 #if CONFIG_MPEG2_CUVID_DECODER
950 case AV_CODEC_ID_MPEG2VIDEO:
951 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG2;
954 #if CONFIG_MPEG4_CUVID_DECODER
955 case AV_CODEC_ID_MPEG4:
956 ctx->cuparseinfo.CodecType = cudaVideoCodec_MPEG4;
959 #if CONFIG_VP8_CUVID_DECODER
960 case AV_CODEC_ID_VP8:
961 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP8;
964 #if CONFIG_VP9_CUVID_DECODER
965 case AV_CODEC_ID_VP9:
966 ctx->cuparseinfo.CodecType = cudaVideoCodec_VP9;
969 #if CONFIG_VC1_CUVID_DECODER
970 case AV_CODEC_ID_VC1:
971 ctx->cuparseinfo.CodecType = cudaVideoCodec_VC1;
975 av_log(avctx, AV_LOG_ERROR, "Invalid CUVID codec!\n");
979 if (avctx->codec->id == AV_CODEC_ID_H264 || avctx->codec->id == AV_CODEC_ID_HEVC) {
980 if (avctx->codec->id == AV_CODEC_ID_H264)
981 bsf = av_bsf_get_by_name("h264_mp4toannexb");
983 bsf = av_bsf_get_by_name("hevc_mp4toannexb");
986 ret = AVERROR_BSF_NOT_FOUND;
989 if (ret = av_bsf_alloc(bsf, &ctx->bsf)) {
992 if (((ret = avcodec_parameters_from_context(ctx->bsf->par_in, avctx)) < 0) || ((ret = av_bsf_init(ctx->bsf)) < 0)) {
993 av_bsf_free(&ctx->bsf);
997 ctx->cuparse_ext.format.seqhdr_data_length = ctx->bsf->par_out->extradata_size;
998 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
999 ctx->bsf->par_out->extradata,
1000 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), ctx->bsf->par_out->extradata_size));
1001 } else if (avctx->extradata_size > 0) {
1002 ctx->cuparse_ext.format.seqhdr_data_length = avctx->extradata_size;
1003 memcpy(ctx->cuparse_ext.raw_seqhdr_data,
1005 FFMIN(sizeof(ctx->cuparse_ext.raw_seqhdr_data), avctx->extradata_size));
1008 ctx->key_frame = av_mallocz(ctx->nb_surfaces * sizeof(int));
1009 if (!ctx->key_frame) {
1010 ret = AVERROR(ENOMEM);
1014 ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
1015 ctx->cuparseinfo.ulMaxDisplayDelay = 4;
1016 ctx->cuparseinfo.pUserData = avctx;
1017 ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
1018 ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
1019 ctx->cuparseinfo.pfnDisplayPicture = cuvid_handle_picture_display;
1021 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1025 ret = cuvid_test_capabilities(avctx, &ctx->cuparseinfo,
1032 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1036 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1037 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1039 if (seq_pkt.payload && seq_pkt.payload_size) {
1040 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1045 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1049 ctx->prev_pts = INT64_MIN;
1051 if (!avctx->pkt_timebase.num || !avctx->pkt_timebase.den)
1052 av_log(avctx, AV_LOG_WARNING, "Invalid pkt_timebase, passing timestamps as-is.\n");
1057 cuvid_decode_end(avctx);
1061 static void cuvid_flush(AVCodecContext *avctx)
1063 CuvidContext *ctx = avctx->priv_data;
1064 AVHWDeviceContext *device_ctx = (AVHWDeviceContext*)ctx->hwdevice->data;
1065 AVCUDADeviceContext *device_hwctx = device_ctx->hwctx;
1066 CUcontext dummy, cuda_ctx = device_hwctx->cuda_ctx;
1067 CUVIDSOURCEDATAPACKET seq_pkt = { 0 };
1070 ret = CHECK_CU(ctx->cudl->cuCtxPushCurrent(cuda_ctx));
1074 av_fifo_freep(&ctx->frame_queue);
1076 ctx->frame_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(CuvidParsedFrame));
1077 if (!ctx->frame_queue) {
1078 av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on flush\n");
1082 if (ctx->cudecoder) {
1083 ctx->cvdl->cuvidDestroyDecoder(ctx->cudecoder);
1084 ctx->cudecoder = NULL;
1087 if (ctx->cuparser) {
1088 ctx->cvdl->cuvidDestroyVideoParser(ctx->cuparser);
1089 ctx->cuparser = NULL;
1092 ret = CHECK_CU(ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo));
1096 seq_pkt.payload = ctx->cuparse_ext.raw_seqhdr_data;
1097 seq_pkt.payload_size = ctx->cuparse_ext.format.seqhdr_data_length;
1099 if (seq_pkt.payload && seq_pkt.payload_size) {
1100 ret = CHECK_CU(ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt));
1105 ret = CHECK_CU(ctx->cudl->cuCtxPopCurrent(&dummy));
1109 ctx->prev_pts = INT64_MIN;
1110 ctx->decoder_flushing = 0;
1114 av_log(avctx, AV_LOG_ERROR, "CUDA reinit on flush failed\n");
1117 #define OFFSET(x) offsetof(CuvidContext, x)
1118 #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1119 static const AVOption options[] = {
1120 { "deint", "Set deinterlacing mode", OFFSET(deint_mode), AV_OPT_TYPE_INT, { .i64 = cudaVideoDeinterlaceMode_Weave }, cudaVideoDeinterlaceMode_Weave, cudaVideoDeinterlaceMode_Adaptive, VD, "deint" },
1121 { "weave", "Weave deinterlacing (do nothing)", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Weave }, 0, 0, VD, "deint" },
1122 { "bob", "Bob deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Bob }, 0, 0, VD, "deint" },
1123 { "adaptive", "Adaptive deinterlacing", 0, AV_OPT_TYPE_CONST, { .i64 = cudaVideoDeinterlaceMode_Adaptive }, 0, 0, VD, "deint" },
1124 { "gpu", "GPU to be used for decoding", OFFSET(cu_gpu), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1125 { "surfaces", "Maximum surfaces to be used for decoding", OFFSET(nb_surfaces), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, INT_MAX, VD },
1126 { "drop_second_field", "Drop second field when deinterlacing", OFFSET(drop_second_field), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VD },
1127 { "crop", "Crop (top)x(bottom)x(left)x(right)", OFFSET(crop_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1128 { "resize", "Resize (width)x(height)", OFFSET(resize_expr), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, VD },
1132 static const AVCodecHWConfigInternal *cuvid_hw_configs[] = {
1133 &(const AVCodecHWConfigInternal) {
1135 .pix_fmt = AV_PIX_FMT_CUDA,
1136 .methods = AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX |
1137 AV_CODEC_HW_CONFIG_METHOD_INTERNAL,
1138 .device_type = AV_HWDEVICE_TYPE_CUDA
1145 #define DEFINE_CUVID_CODEC(x, X) \
1146 static const AVClass x##_cuvid_class = { \
1147 .class_name = #x "_cuvid", \
1148 .item_name = av_default_item_name, \
1149 .option = options, \
1150 .version = LIBAVUTIL_VERSION_INT, \
1152 AVCodec ff_##x##_cuvid_decoder = { \
1153 .name = #x "_cuvid", \
1154 .long_name = NULL_IF_CONFIG_SMALL("Nvidia CUVID " #X " decoder"), \
1155 .type = AVMEDIA_TYPE_VIDEO, \
1156 .id = AV_CODEC_ID_##X, \
1157 .priv_data_size = sizeof(CuvidContext), \
1158 .priv_class = &x##_cuvid_class, \
1159 .init = cuvid_decode_init, \
1160 .close = cuvid_decode_end, \
1161 .decode = cuvid_decode_frame, \
1162 .receive_frame = cuvid_output_frame, \
1163 .flush = cuvid_flush, \
1164 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_AVOID_PROBING | AV_CODEC_CAP_HARDWARE, \
1165 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_CUDA, \
1169 AV_PIX_FMT_NONE }, \
1170 .hw_configs = cuvid_hw_configs, \
1171 .wrapper_name = "cuvid", \
1174 #if CONFIG_HEVC_CUVID_DECODER
1175 DEFINE_CUVID_CODEC(hevc, HEVC)
1178 #if CONFIG_H264_CUVID_DECODER
1179 DEFINE_CUVID_CODEC(h264, H264)
1182 #if CONFIG_MJPEG_CUVID_DECODER
1183 DEFINE_CUVID_CODEC(mjpeg, MJPEG)
1186 #if CONFIG_MPEG1_CUVID_DECODER
1187 DEFINE_CUVID_CODEC(mpeg1, MPEG1VIDEO)
1190 #if CONFIG_MPEG2_CUVID_DECODER
1191 DEFINE_CUVID_CODEC(mpeg2, MPEG2VIDEO)
1194 #if CONFIG_MPEG4_CUVID_DECODER
1195 DEFINE_CUVID_CODEC(mpeg4, MPEG4)
1198 #if CONFIG_VP8_CUVID_DECODER
1199 DEFINE_CUVID_CODEC(vp8, VP8)
1202 #if CONFIG_VP9_CUVID_DECODER
1203 DEFINE_CUVID_CODEC(vp9, VP9)
1206 #if CONFIG_VC1_CUVID_DECODER
1207 DEFINE_CUVID_CODEC(vc1, VC1)